Step 1 β Data ingestion. Download the authors' reconstruction files (.swc) and analysis code from Zenodo (doi:10.5281/zenodo.22801634 and 10.5281/zenodo.22808315) (). Parse with Python (morphio or neurom) into branching graphs; map each bifurcation node to inter-bifurcation path distance, branch order, and parent-child angles.
Step 2 β Replicate core statistics. Recompute: bifurcation enrichment (47/58 swellings at bifurcations, ~81%, 95% CI 0.69β0.89, one-sided binomial p < 0.001 vs sliding-window chance baseline); axonal identity (55/58 swellings, 94.8% MBP-positive); soma survival (99.9% controls vs 89.4% P301S at 3 months, exact permutation Mann-Whitney p = 0.0022, N=6/group) ().
Step 3 β Geometric risk model. Refit the logistic regression on inter-bifurcation distance with leave-one-neuron-out CV; verify LOO AUC 0.852 and Poisson-binomial risk-map tagging (top-1 in 7/10 neurons, p = 0.0004). Test angle features (weak: LOO AUC 0.574/0.555) as model checks ().
Step 4 β Extension/falsification. Apply the LOO model to held-out arbors: AUC near chance or failure to tag swelling sites would weaken the central claim. Test whether local branch order or diameter better explains vulnerability than distance alone.
Known limits: AAV-P301S is accelerated and cell-autonomous, likely exceeding endogenous tau burden; only 10/37 patched neurons were reconstructed; molecular mechanism of branch vulnerability is unresolved; human CBD neuropil tau lies mostly in non-PV cells ().
Know what changed, what holds up, and what remains uncertain. Every Friday. No ads.